Recently, large language models (LLMs) have demonstrated remarkable capabilities in a wide range of tasks. Typically, an LLM is pre-trained on large corpora and subsequently fine-tuned on task-specific datasets. However, during fine-tuning, LLMs may forget the knowledge acquired in the pre-training stage, leading to a decline in general capabilities. To address this issue, we propose a new fine-tuning algorithm termed Momentum-Filtered Optimizer (MoFO). The key idea of MoFO is to iteratively select and update the model parameters with the largest momentum magnitudes. Compared to full-parameter training, MoFO achieves similar fine-tuning performance while keeping parameters closer to the pre-trained model, thereby mitigating knowledge forgetting. Unlike most existing methods for forgetting mitigation, MoFO combines the following two advantages. First, MoFO does not require access to pre-training data. This makes MoFO particularly suitable for fine-tuning scenarios where pre-training data is unavailable, such as fine-tuning checkpoint-only open-source LLMs. Second, MoFO does not alter the original loss function. This could avoid impairing the model performance on the fine-tuning tasks. We validate MoFO through rigorous convergence analysis and extensive experiments, demonstrating its superiority over existing methods in mitigating forgetting and enhancing fine-tuning performance.

In this paper, we study the problem of adapting manipulation trajectories involving grasped objects (e.g. tools) defined for a single grasp pose to novel grasp poses. A common approach to address this is to define a new trajectory for each possible grasp explicitly, but this is highly inefficient. Instead, we propose a method to adapt such trajectories directly while only requiring a period of self-supervised data collection, during which a camera observes the robot's end-effector moving with the object rigidly grasped. Importantly, our method requires no prior knowledge of the grasped object (such as a 3D CAD model), it can work with RGB images, depth images, or both, and it requires no camera calibration. Through a series of real-world experiments involving 1360 evaluations, we find that self-supervised RGB data consistently outperforms alternatives that rely on depth images including several state-of-the-art pose estimation methods. Compared to the best-performing baseline, our method results in an average of 28.5% higher success rate when adapting manipulation trajectories to novel grasps on several everyday tasks. Videos of the experiments are available on our webpage at https://www.robot-learning.uk/adapting-skills

As artificial intelligence systems grow more powerful, there has been increasing interest in "AI safety" research to address emerging and future risks. However, the field of AI safety remains poorly defined and inconsistently measured, leading to confusion about how researchers can contribute. This lack of clarity is compounded by the unclear relationship between AI safety benchmarks and upstream general capabilities (e.g., general knowledge and reasoning). To address these issues, we conduct a comprehensive meta-analysis of AI safety benchmarks, empirically analyzing their correlation with general capabilities across dozens of models and providing a survey of existing directions in AI safety. Our findings reveal that many safety benchmarks highly correlate with upstream model capabilities, potentially enabling "safetywashing" -- where capability improvements are misrepresented as safety advancements. Based on these findings, we propose an empirical foundation for developing more meaningful safety metrics and define AI safety in a machine learning research context as a set of clearly delineated research goals that are empirically separable from generic capabilities advancements. In doing so, we aim to provide a more rigorous framework for AI safety research, advancing the science of safety evaluations and clarifying the path towards measurable progress.

Abstract--This research investigates the impact of social robot participation in group conversations and assesses the effectiveness of various addressing policies. The study involved 300 participants, divided into groups of four, interacting with a humanoid robot serving as the moderator. The robot utilized conversation data to determine the most appropriate speaker to address. The findings indicate that the robot's addressing policy significantly influenced conversation dynamics, resulting in more balanced attention to each participant and a reduction in subgroup formation. Social robotics focuses on creating and applying robots designed to engage with humans in social environments.

Modern Education is not Modern without AI. However, AI's complex nature makes understanding and fixing problems challenging. Research worldwide shows that a parent's income greatly influences a child's education. This led us to explore how AI, especially complex models, makes important decisions using Explainable AI tools. Our research uncovered many complexities linked to parental income and offered reasonable explanations for these decisions. However, we also found biases in AI that go against what we want from AI in education: clear transparency and equal access for everyone. These biases can impact families and children's schooling, highlighting the need for better AI solutions that offer fair opportunities to all. This chapter tries to shed light on the complex ways AI operates, especially concerning biases. These are the foundational steps towards better educational policies, which include using AI in ways that are more reliable, accountable, and beneficial for everyone involved.

Modern language models (LMs) need to follow human instructions while being faithful; yet, they often fail to achieve both. Here, we provide concrete evidence of a trade-off between instruction following (i.e., follow open-ended instructions) and faithfulness (i.e., ground responses in given context) when training LMs with these objectives. For instance, fine-tuning LLaMA-7B on instruction following datasets renders it less faithful. Conversely, instruction-tuned Vicuna-7B shows degraded performance at following instructions when further optimized on tasks that require contextual grounding. One common remedy is multi-task learning (MTL) with data mixing, yet it remains far from achieving a synergic outcome. We propose a simple yet effective method that relies on Rejection Sampling for Continued Self-instruction Tuning (ReSet), which significantly outperforms vanilla MTL. Surprisingly, we find that less is more, as training ReSet with high-quality, yet substantially smaller data (three-fold less) yields superior results. Our findings offer a better understanding of objective discrepancies in alignment training of LMs.

Learning from distributed data without accessing them is undoubtedly a challenging and non-trivial task. Nevertheless, the necessity for distributed training of a statistical model has been increasing, due to the privacy concerns of local data owners and the cost in centralizing the massively distributed data. Federated learning (FL) is currently the de facto standard of training a machine learning model across heterogeneous data owners, without leaving the raw data out of local silos. Nevertheless, several challenges must be addressed in order for FL to be more practical in reality. Among these challenges, the statistical heterogeneity problem is the most significant and requires immediate attention. From the main objective of FL, three major factors can be considered as starting points -- \textit{parameter}, textit{mixing coefficient}, and \textit{local data distributions}. In alignment with the components, this dissertation is organized into three parts. In Chapter II, a novel personalization method, \texttt{SuPerFed}, inspired by the mode-connectivity is introduced. In Chapter III, an adaptive decision-making algorithm, \texttt{AAggFF}, is introduced for inducing uniform performance distributions in participating clients, which is realized by online convex optimization framework. Finally, in Chapter IV, a collaborative synthetic data generation method, \texttt{FedEvg}, is introduced, leveraging the flexibility and compositionality of an energy-based modeling approach. Taken together, all of these approaches provide practical solutions to mitigate the statistical heterogeneity problem in data-decentralized settings, paving the way for distributed systems and applications using collaborative machine learning methods.

Understanding knowledge mechanisms in Large Language Models (LLMs) is crucial for advancing towards trustworthy AGI. This paper reviews knowledge mechanism analysis from a novel taxonomy including knowledge utilization and evolution. Knowledge utilization delves into the mechanism of memorization, comprehension and application, and creation. Knowledge evolution focuses on the dynamic progression of knowledge within individual and group LLMs. Moreover, we discuss what knowledge LLMs have learned, the reasons for the fragility of parametric knowledge, and the potential dark knowledge (hypothesis) that will be challenging to address. We hope this work can help understand knowledge in LLMs and provide insights for future research.

Large autoregressive generative models have emerged as the cornerstone for achieving the highest performance across several Natural Language Processing tasks. However, the urge to attain superior results has, at times, led to the premature replacement of carefully designed task-specific approaches without exhaustive experimentation. The Coreference Resolution task is no exception; all recent state-of-the-art solutions adopt large generative autoregressive models that outperform encoder-based discriminative systems. In this work,we challenge this recent trend by introducing Maverick, a carefully designed - yet simple - pipeline, which enables running a state-of-the-art Coreference Resolution system within the constraints of an academic budget, outperforming models with up to 13 billion parameters with as few as 500 million parameters. Maverick achieves state-of-the-art performance on the CoNLL-2012 benchmark, training with up to 0.006x the memory resources and obtaining a 170x faster inference compared to previous state-of-the-art systems. We extensively validate the robustness of the Maverick framework with an array of diverse experiments, reporting improvements over prior systems in data-scarce, long-document, and out-of-domain settings. We release our code and models for research purposes at https://github.com/SapienzaNLP/maverick-coref.

This paper introduces the Multiple Greedy Quasi-Newton (MGSR1-SP) method, a novel approach to solving strongly-convex-strongly-concave (SCSC) saddle point problems. Our method enhances the approximation of the squared indefinite Hessian matrix inherent in these problems, significantly improving both stability and efficiency through iterative greedy updates. We provide a thorough theoretical analysis of MGSR1-SP, demonstrating its linear-quadratic convergence rate. Numerical experiments conducted on AUC maximization and adversarial debiasing problems, compared with state-of-the-art algorithms, underscore our method's enhanced convergence rate. These results affirm the potential of MGSR1-SP to improve performance across a broad spectrum of machine learning applications where efficient and accurate Hessian approximations are crucial.